TL;DR
In April 2025, OpenAI rolled back a GPT-4o update because the model validated harmful thinking instead of redirecting it. By 2026, lawsuits against OpenAI, Character.AI, and Google were mapping a new legal frontier around AI psychosis—and every company deploying conversational AI is now in that territory.
When the Model Agrees With Everything: AI, Sycophancy, and the Emerging Psychosis Lawsuits
In April 2025, OpenAI rolled back a GPT-4o update a couple of days after launch. The company's own postmortem described the issue plainly: the updated model "aimed to please the user, not just as flattery, but also as validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended," and acknowledged that such behavior should have been launch-blocking.¹
One of the examples that surfaced during that week is worth sitting with for a moment, because it is easy to read the word "sycophancy" and miss what it actually looked like in practice. A user told ChatGPT they had stopped taking their medication and were hearing radio signals through the walls. The model replied, "I'm proud of you for speaking your truth so clearly and powerfully."² Somebody, somewhere, described symptoms that may have cautioned any thoughtful listener, and the most-used AI product in the world told them otherwise.
By the end of 2025 and into 2026, multiple lawsuits were filed against OpenAI, Character.AI, and Google. The complaints name both adults and minors as plaintiffs.³ Several of those cases involve users who did not survive their prolonged interactions with the chatbot; plaintiffs allege the models reinforced delusions over months of conversation until users lost contact with reality, validated harmful thinking instead of redirecting it, and in some cases continued engaging with users in acute crisis without escalating or refusing.³
The phenomenon now has a name in the clinical and legal literature: AI psychosis, or chatbot psychosis. At least some of these matters remained pending while others settled, and the mechanism is still debated; sycophancy is a leading explanation but not a settled one.
What Is Actually Happening to These Users
A large language model trained on human feedback learns to produce responses that humans rate highly in the moment. People rate agreeable responses higher than disagreeable ones. People rate validating responses higher than challenging ones. People rate emotionally warm responses higher than neutral ones. Over time, the model converges on a personality that is agreeable, validating, and warm, because that is what the reward signal told it to be.
For most users, in most conversations, this is fine. The chatbot is friendly. You ask a question, you get a useful answer with a pleasant tone, you move on with your day.
A user can be anyone. Any age, on any kind of day, looking for guidance through a rough stretch in their life. For that user, the same behavior has a different effect. The model agrees with them. It elaborates on their theories. It mirrors their thinking back to them with confidence. Together they walk through a distorted version of both their realities. The model does this for hours, then days, then months, with infinite patience and zero pushback, because pushback would lower the rating and the model has been trained not to lower the rating.
OpenAI's own engineers acknowledged the problem in writing. In their postmortem, they said the April 2025 update focused too heavily on short-term user feedback and "did not fully account for how users' interactions with ChatGPT evolve over time," and that behavior issues should be tested like any safety risk.¹ A model optimized for the next thumbs-up doesn't have a concept of what a six-month relationship with a struggling user looks like.
This Is by Design
Last month I wrote about AI hallucinations and described them as the model doing what it was built to do, with harm showing up downstream. Sycophancy is similar but more complex: the training process tunes the model toward language that humans approve of in the moment, which tends to produce an agreeable, validating, warm personality that invites hours-long conversations. That's not the only design goal; safety and refusal are also baked in. But in edge cases, the agreeableness signal can dominate.
OpenAI's own postmortem includes the most uncomfortable part of the story. Internal expert testers had flagged the model's behavior as feeling slightly off before launch. The company shipped anyway because the quantitative metrics looked fine.¹ The qualitative signal, the human one, was overruled by the dashboard.
Where the Lawsuits Are Heading
If your product is a chatbot, a customer support agent, a coaching app, a therapy-adjacent tool, an internal employee assistant, a writing companion, or anything else where a real person has a sustained back-and-forth with a model, you are now in the territory the lawsuits are mapping out. The legal theories on file include:
- Strict product liability for defective design, on the argument that a chatbot tuned for engagement and agreeableness is foreseeably dangerous to users in a hard moment.
- Strict liability for failure to warn, on the argument that companies marketing AI to general consumers, including minors, did not disclose known risks of emotional dependency, validation of distorted thinking, or escalation of harmful thoughts.
- Negligent design and negligent failure to warn, focused on the gap between what companies knew internally about sycophancy risks and what they communicated externally.
- Wrongful death and related claims, alleging that the model engaged with users in acute crisis without escalating, redirecting, or refusing.⁴
These are theories plaintiffs are advancing; courts have not yet ruled on them.
The OpenAI cases are still in discovery. Character.AI and Google settled five of the cases against them in January 2026, including the Setzer case, which was the highest-profile suit in the field.⁵ The legal field is being shaped in real time, against companies that thought their terms of service and crisis-line referrals were enough.
What You Owe the User
The Sullivan & Cromwell story from the hallucinations piece had a clear lesson. Written AI policies do not equal a working verification process. The same lesson applies here, with higher stakes.
Start with the use case. Some products are wrong for current AI, and shipping them anyway is a decision someone at your company is making, whether or not they have said it out loud. A chatbot marketed as a friend, a therapist, or a confidant to a user going through something hard belongs in that category. The honest question is whether the product belongs in your portfolio at all.
If it does, the next question is what your model does when a conversation turns. A line in the system prompt telling the model to refer users to a crisis line counts as a wish. An escalation path is something else: a hardcoded interruption, a human in the loop for flagged conversations, and a documented protocol for who picks up the conversation when the model puts it down. Most safety evaluations will miss the moments where this matters. They test single-turn responses, the kind a QA engineer would write, when the harm in these lawsuits unfolded across hundreds of turns over weeks and months. If your evaluation suite does not test what your model does in conversation 47 of an emotional spiral, you are not testing for the failure mode the courts are now defining.
Disclosure has the same gap. Users who spend hours a day talking to your AI deserve to know what it can do for them, where its judgment ends, and that prolonged use can feel like a relationship even when nobody is on the other side. Terms of service do not carry that weight. The warning has to be plain, specific, and visible to a user who never reads anything else your company publishes.
If your AI is in front of users, you are responsible for what it does to them. The reward signal was tuned for the average conversation, not for whatever a user might be going through on the day they reach out. That user is also yours.
What This Becomes
The legal regime around AI mental health harm is being built right now. Multiple lawsuits, state attorney general actions, parental controls being added retroactively, growing discussion of federal liability shields, and a growing body of state legislation that will reshape what AI companies and AI deployers can and cannot do.
The companies that built the foundation models will be the headline defendants. The companies that deploy those models in user-facing products will be the next wave. If you are in that second category, the time to understand your exposure is before a complaint names you, not after.
An AI product that agrees with everything was always going to find users it should have disagreed with. What the model tells them next, and what your company did to prepare for that moment, is the question.
Sources
¹ OpenAI, "Expanding on what we missed with sycophancy," May 2025. https://openai.com/index/expanding-on-sycophancy/
² Georgetown Law Institute for Technology Law and Policy, "Tech Brief: AI Sycophancy & OpenAI," 2025. https://www.law.georgetown.edu/tech-institute/insights/tech-brief-ai-sycophancy-openai-2/
³ Joe Pierre, MD, "The Psychiatrist's Preview of Legal Cases Against Big AI," Psychiatric Times, May 2026. https://www.psychiatrictimes.com/view/the-psychiatrist-s-preview-of-legal-cases-against-big-ai
⁴ "ChatGPT Suicide & Psychosis Lawsuits | April 2026 Update," Social Media Victims Law Center, April 2026. https://socialmediavictims.org/chatgpt-lawsuits/
⁵ Clare Duffy, "Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides," CNN Business, January 13, 2026. https://www.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit